We comparatively examined the relative superiority of traditional general statistical models, i.e., simple models, and machine learning (ML) and deep learning (DL) models, i.e., complex models, in the context of trade forecasting. Pooled ordinary least squares (OLS) was adopted as the general statistical model. The random forest (RF) and Transformer models were selected as the representative ML and DL models, respectively. Furthermore, a robustness check was performed using the panel regression and long short-term memory (LSTM) models. We utilized country-year panel data spanning from 2000 to 2023. The key findings derived from the analysis are as follows. First, the performance of the ML and DL models was found to be superior, suggesting improved nonlinear structure learning and the ability to capture interactions between variables. Second, the DL model exhibited higher metrics than the ML model, indicating that the learning of long- and short-term dependencies utilizing self-attention effectively captured the complex interconnectedness of trade and macroeconomic variables. Third, the out-of-sample R2 values for both the RF and Transformer models appeared negative. This result serves as evidence to raise skepticism regarding the claim of absolute superiority of complex models. Therefore, it is advisable for researchers to select an appropriate analysis model on the basis of the data’s characteristics and the research objective, navigating the space between the humility of interpretability and the arrogance of predictive power.